Architecture-Level Event-Oriented Frontend for Heterogeneous Sensor Inference in Photonic Stochastic Systems

Heterogeneous sensor systems generate measurements in incompatible physical units, which complicates their direct integration with photonic stochastic processors. This study proposes an architecture-level event-oriented frontend that converts heterogeneous sensor channels into unified event-oriented probabilities and then into Bernoulli bitstreams compatible with polarization-encoded optical interfaces. The framework combines sensor-to-probability mapping, weighted event-level fusion, stochastic bitstream generation, and system-level control of correlation and synchronization. Its performance was investigated through reproducible Colab-based modeling using baseline validation, weighting-strategy comparison, static and time-varying decorrelation/synchronization studies, and robustness/scaling analysis. The validation framework also includes a first-order analysis of photonic-interface nonidealities and a semi-real validation case using a digitized 520 nm optical smoke-transmittance response from the literature. The results show that the stochastic event estimate converges toward the float reference with increasing bitstream length, reliability-aware weighting outperforms equal and tested data-driven weighting in the benchmark, independent stream generation provides the best inference quality, and synchronization mismatch becomes measurable in time-varying fusion. The frontend also demonstrates graceful degradation under channel corruption and favorable scaling under mixed informative, weak, redundant, conflicting, and noisy channel configurations. These findings indicate that heterogeneous sensors can be interfaced with photonic stochastic systems through a common event-level representation and that weighting, decorrelation, synchronization, and robustness must be treated as core frontend design variables.